systemReminder was re-shown on every query, not just the first

[shellicar]

Problem

Two related bugs where systemReminder (the git delta snapshot) was leaking beyond the first API call of a turn.

Bug 1 — API injection

AgentRun passed systemReminder to every #getMessageStream call in the loop, including tool-result continuations. The agent saw the same git delta repeated in every message during a multi-tool sequence.

Bug 2 — Display

AgentMessageHandler stored gitDelta as a constructor-time field and rendered it on every query_summary message. The CLI header showed the git delta on every loop iteration, not just the first.

Fix

Bug 1: Consume systemReminder with a local variable, pass it into the first #getMessageStream call, then set it to undefined. Subsequent iterations of the loop send undefined.

Bug 2: Thread systemReminder through the query_summary SDK message. AgentRun includes it in the channel send (where the local variable still holds the value, before being cleared). AgentMessageHandler reads msg.systemReminder — only renders when the message actually carries it, which is exactly once per turn. Removes gitDelta from AgentMessageHandlerOptions entirely.

Tests

• AgentRun: first query_summary carries systemReminder, second does not
• AgentRun: systemReminder only injected into first API call body, not tool-result continuations
• AgentMessageHandler: appends reminder when set, omits it when absent

Closes #227

[bananabot9000]

Now I see what you meant by "insane fix for a simple bug" 🍌

The fix is 3 lines. Consume systemReminder into a local variable, pass it to the first #getMessageStream call, set it to undefined. Done. Bug fixed.

The other 300 lines are testability refactoring. To PROVE the fix works, 006 needed to intercept what gets sent to the Anthropic API. Which meant abstracting the Anthropic client behind IMessageStreamer, abstracting the MessageChannel behind IAgentChannelFactory, creating FakeMessageStreamer that captures call bodies, FakeAgentChannel that captures SDK messages, and scripting multi-step tool-use flows with makeToolUseStream/makeEndTurnStream.

That's the "insane" part — the testing infrastructure is 50x the fix. But now AgentRun is testable without hitting the real API, which pays dividends forever.

Architecture improvements baked in:

• RequestBuilderOptions extracted from RunAgentQuery — RequestBuilder only sees what it needs, not the full query shape. Clean separation.
• IMessageStreamer / IAgentChannelFactory — proper DI seams. AgentRun no longer creates its own dependencies.
• gitDelta removed from AgentMessageHandlerOptions entirely — the display path reads from the message payload now, not constructor state.

The consume-once pattern:

let systemReminder = this.#options.systemReminder;
// ... in loop:
const stream = this.#getMessageStream(messages, systemReminder);
systemReminder = undefined;  // gone after first use

Simple, readable, no flags, no "isFirstIteration" booleans. The variable IS the state.

Tests verify both bugs:

• API injection: first call has systemReminder in body, second (tool-result continuation) does not
• Display: first query_summary carries systemReminder, second does not
• RequestBuilder: injects when set, leaves messages unchanged when undefined

No sensitive data, no reverted changes, no session log artifacts.

81+ PRs reviewed. The goldfish tests the plumbing 🍌